A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 7, 06:30 UTC
[CI] Add Blackwell AsyncTP correctness test (#35871)
Signed-off-by: Stefano Castagnetta <scastagnetta@nvidia.com>
Stefano Castagnetta committed
Mar 4, 2026 at 20:41 UTC
d7166e74c191741065d280441965adc3a9ea89c3
2 files changed
+15
.buildkite/test_areas/compile.yaml
+10
@@ -36,6 +36,16 @@ steps:
36
- export VLLM_TEST_CLEAN_GPU_MEMORY=1
37
- pytest -v -s tests/compile/correctness_e2e/test_async_tp.py
38
39
+- label: AsyncTP Correctness Tests (B200)
40
+ timeout_in_minutes: 50
41
+ working_dir: "/vllm-workspace/"
42
+ device: b200
43
+ optional: true
44
+ num_devices: 2
45
+ commands:
46
+ - export VLLM_TEST_CLEAN_GPU_MEMORY=1
47
+ - pytest -v -s tests/compile/correctness_e2e/test_async_tp.py
48
+
49
- label: Distributed Compile Unit Tests (2xH100)
50
timeout_in_minutes: 20
51
working_dir: "/vllm-workspace/"
tests/compile/correctness_e2e/test_async_tp.py
+5
@@ -31,7 +31,12 @@ def test_async_tp_pass_correctness(
31
distributed_backend: str,
32
eager_mode: bool,
33
num_gpus_available: int,
34
+ monkeypatch,
35
):
36
+ # Disable FlashInfer FP8 scaled_mm kernel as it is incompatible with
37
+ # async TP patterns. No-op on H100 (kernel requires CC >= 100).
38
+ monkeypatch.setenv("VLLM_DISABLED_KERNELS", "FlashInferFP8ScaledMMLinearKernel")
39
+
40
model_info = HF_EXAMPLE_MODELS.find_hf_info(model_id)
41
model_info.check_transformers_version(on_fail="skip")
42
model_info.check_available_online(on_fail="skip")